Intro to Machine Learning Classification with Python

How to get started with classification whether you're a developer or an aspiring data scientist. Start by breaking the problem down by asking the right questions.

Transcript

1. Intro to Machine Learning Classification with Python @bigdana
2. Objectives @bigdana classification vs regression navigate better - use an api or scikit-learn? multi-class vs multi-label classification performance evaluation down sampling over fitting
3. Mode of Transport @bigdana
4. Step 1: Do I have a problem appropriate for classification? @bigdana Can I make one?
5. The Data @bigdana 9 Democratic Debates 12 Republican Debates annotated with: “applause” “cheering” “laughter” “booing” source: http://www.presidency.ucsb.edu/debates.php total words: 56118 Hillary Donald total words: 40699 total turns: 564 total turns: 774
6. Binary Classification @bigdana Hillary Donald “I think the big problem this country has is being politically correct.”
7. Binary Classification @bigdana No Applause Applause “Oh, for goodness — that's not going to happen. I'm not even answering that question.”
8. multi-class classification @bigdana positive controversial neutral negative “First of all, there's nobody on this stage that's more pro Israel than I am. OK. There's nobody. I am pro-Israel.” laughter, cheering applause OR (laughter, cheering applause) OR AND booing booing
9. Multi-label Classification @bigdana
10. @bigdana IBM Watson: http://watson-personality-insights-python.mybluemix.net/ Multi-label Classification
11. @bigdana Multi-label Classification IBM Watson: http://watson-personality-insights-python.mybluemix.net/
12. maybe classification isn’t for you @bigdana
13. Regression @bigdana Price Weight Height Temperature Distance
14. Classification vs Regression @bigdana
15. Who are you? @bigdana
16. Developer @bigdana
17. Try a Machine Learning API @bigdana IBM Watson Azure Machine Learning API AWS API BIGML Google Prediction API
18. Pros and Cons of Google Prediction API @bigdana pros: hosted model stream training data cons: standard rate limit: 1 prediction per second confusion matrix is not super intuitive to evaluate performance
19. What you’ll need to know/learn @bigdana pandas feature selection how to work with an api - oauth credentials how to interpret a confusion matrix
20. Sample Confusion Matrix @bigdana Binary Classifier
21. Sample Confusion Matrix @bigdana Multi-Class Classifier
22. (aspiring) Data Scientist @bigdana
23. I want to work with prebuilt models @bigdana Scikit-learn models: Logistic Regression Support Vector Machines Stochastic Gradient Descent Nearest Neighbors Decision Trees Ensemble Methods: Gradient Tree Boosting
24. What you will need to know/learn @bigdana pandas feature selection down sampling which models to play with and how how to interpret precision, recall and f1 scores for performance evaluation
25. I want to learn how to build my own models! @bigdana Machine Learning by Andrew Ng on Coursera: Tensor Flow: https://github.com/aymericdamien/TensorFlow-Examples https://www.coursera.org/learn/machine-learning Hybrid approach: Scikit-Flow
26. What you will need to know/learn @bigdana matrix algebra calculus Machine Learning by Andrew Ng on Coursera: Tensor Flow: A machine learning course?
27. Machine Learning difficulty @bigdana a machine learning API Scikit Learn Tensor Flow
28. Machine Learning difficulty @bigdana a machine learning API Scikit Learn Tensor Flow
29. Let’s give it a go! @bigdana
30. 80/20 @bigdana Training set Test set
31. 80/20 @bigdana Training set Test set X y
32. 80/20 @bigdana Training set Test set Well, first of all, it's great to be here in New York, and I am delighted to...... have this chance to discuss the issues that are important to our future. I was so honored to serve as a Well, it is true that now that the spotlight is pretty bright here in New York, some things have been said and Senator Sanders did call me unqualified. I've been called a Well, let me... ...let me just say... ...let me...let me say... 1 1 0 0 0 X y Clinton
33. 80/20 @bigdana Training set Test set I fully understand. I cannot say. I have to respect the person that, if it's not me, the person that wins, if I do win, and I'm leading by quite a bit, that's what I want to do. I can totally make that pledge. If Well, I've given him plenty of money. I will not make the pledge at this time. Only Rosie O'Donnell. 0 0 0 0 0 X y Trump
34. @bigdana
35. Excellent Precision @bigdana If you say it gets an applause - it does. You might miss some that should get an applause.
36. Excellent Recall @bigdana You are really good at catching all the applauses! You sometimes catch a lot that aren’t applauses though.
37. F1 score = Harmonic Mean @bigdana
38. @bigdana
39. @bigdana
40. Down Sampling @bigdana
41. Down Sampling @bigdana More balanced!
42. Other things to play with: @bigdana Feature Selection Try different models!
43. @bigdana
44. Corrections? Suggestions? Questions? @bigdana Thanks!